How Deep is the Inside Story of DeepSeek? Model Source Sparks Controversy

China’s artificial intelligence startup DeepSeek is currently attracting domestic and international attention, with some media touting it as challenging the leading position of American AI.

However, in China where information lacks public transparency, mysteries still shroud DeepSeek’s specific financial, material, and manpower investments behind its technical breakthroughs. There are even speculations of potential involvement in illegal data theft from OpenAI. Analysts believe that DeepSeek cannot shake off the hat of self-censorship under the control of the Chinese Communist Party, as international observers have noticed its filtering of sensitive words related to the CCP during testing.

On Monday, January 27, chip manufacturer Nvidia saw its stock price drop by 17%, causing Nasdaq to open with a 3% decline. Western media analyzed that this was due to DeepSeek’s release of the new AI model R1 on January 20, sparking concerns about future demand for Nvidia products and market share, leading to the second-largest single-day stock price drop in Nvidia’s history.

DeepSeek is a product of Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. (referred to as “DeepSeek Company”). Established in July 2023, by China’s quant giant High-Flyer (Hangzhou High-Flyer Technology Co., Ltd.), the 40-year-old Liang Wenfeng is the founder of both companies and also the CEO of DeepSeek.

When DeepSeek unveiled its new model R1 on January 20, Liang Wenfeng met with Chinese Premier Li Keqiang and Vice Premier Ding Xuexiang. Li Keqiang presided over a small seminar with experts, entrepreneurs, and representatives from various fields to hear suggestions on government work reports, where Liang Wenfeng was one of the nine speakers. Li Keqiang emphasized the need to “fully implement the decisions and arrangements of the Party Central Committee.” Liang Wenfeng was prominently featured on CCTV’s “News Broadcast.”

Last year, a similar conference was attended by Baidu’s co-founder, chairman, and CEO Li Yanhong. In March 2023, Baidu officially launched its dialogue-based AI, Ernie Bot, based on large-scale pre-trained language models, taking a significant step into the generative AI field and competing with international models like ChatGPT.

DeepSeek has been observed to self-censor sensitive topics related to the CCP. For example, when asked about topics like “Possibility of CCP invading Taiwan” or “Tiananmen Square incident,” or even the name “Xi Jinping,” DeepSeek would censor the conversation or divert it to other topics, avoiding giving straightforward replies.

Since August 15, 2023, China has implemented the “Interim Measures for the Management of Generative Artificial Intelligence Services,” requiring adherence to socialist core values and prohibiting content that endangers national security or unity. These regulations necessitate filtering out sensitive words and topics by services like DeepSeek.

This policy document, approved by the National Internet Information Office of the CCP, along with various ministries and bureaus, places restrictions on generative AI services, including content adjustments to align with socialist values.

DeepSeek stands out for its emphasis on the security of responses. In January 2024, they published a paper titled “DeepSeek LLM: Expanding Open Source Language Models with Long Termism,” highlighting their commitment to ensuring model security throughout the training process. They formed a 20-member team of experts from various fields to verify the model’s security, along with establishing a content classification system.

Apart from security content diversity, DeepSeek focuses on the format diversity of secure content. They paid attention to constructing a variety of secure questions by inducing, role-playing, multi-turn conversations, and default positions to avoid providing unsafe responses based on surface formats.

Only after passing tests and audits by the Central Cyberspace Administration, can domestic AI large models in China go online. As of December 30, 2024, 302 Chinese AI large models have been registered with the National Cyberspace Administration of China.

According to official announcements, DeepSeek’s registered large model is called “Seek Chat DeepSeek Chat,” while the registered algorithm is named “DeepSeek Large Language Model Algorithm,” both under Beijing DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.

It’s worth noting that the Beijing-based company has received limited media attention compared to reports focusing on DeepSeek and its relationship with Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. The Hangzhou-based company, as a service provider, registered an algorithm called “DeepSeekChat Seek Dialogue Generation Algorithm” for application on DeepSeek’s website.

In July 2024, Chinese media 36Kr published an exclusive interview with Liang Wenfeng, where he mentioned the challenges they face due to restrictions on high-end chips due to sanctions.

Reports have circulated about DeepSeek’s new generation open-source large model, DeepSeek-v3, released in December 2024, claiming that it only requires 2,024 Nvidia H800 chips for training at a cost of 5.58 million USD. The original source of this information remains unconfirmed by Epoch Times reporters.

The company behind DeepSeek, High-Flyer, had reportedly accumulated tens of thousands of Nvidia chips, evident in Liang Wenfeng’s involvement in the paper: “In deep learning training, we deployed the Firefly II system equipped with ten thousand A100 GPU chips.” High-Flyer launched a deep learning training platform “Firefly II” in 2021, with an investment of about 1 billion CNY. They claim that the machine was used for scientific research support and idle computing power reached 15.33 million GPU hours in 2020.

Despite the exact amount of resources invested in DeepSeek remaining undisclosed, the company is openly competing with OpenAI with a focus on affordability. Their latest versions, DeepSeek-R1 and DeepSeek-V3, are claimed to match the performance of leading closed-source models.

On Wednesday, January 29, an OpenAI spokesperson acknowledged indications suggesting DeepSeek had inappropriately distilled OpenAI models. In response, OpenAI is actively investigating and taking measures to protect their technology, emphasizing the importance of adhering to service terms that prohibit replicating their models.

“Distillation” is a developmental technique involving teaching smaller models using knowledge from larger and more advanced models to achieve similar performance on specific tasks.

Jason, an AI expert, expressed concerns about the security of information collected by DeepSeek. While user agreements disclose some information collected, potential unauthorized use of user data in the background poses risks.

Jason stated, “Given DeepSeek’s location in China, it operates under Chinese regulations. If the CCP demands actions from DeepSeek, it could be challenging to refuse.” The potential pressure and surveillance from the Chinese government could compromise user data security.

In light of prominent figures using DeepSeek, Jason believes the Chinese government would unlikely allow DeepSeek to operate independently, especially with laws like the “Anti-Espionage Law” and the “National Intelligence Law” facilitating surveillance.

The long-term sustainability of DeepSeek’s open-source support is questioned by Miles Brundage, an independent AI policy researcher, highlighting concerns about possible backdoors in models related to the Chinese government. He believes that the worries about China leading the open-source domain are valid from an American perspective.

Concerns over the security of information collected by DeepSeek persist. The platform discloses collecting various personal data, including user-provided information and automatically collected data, raising privacy and surveillance concerns.

DeepSeek’s data may be stored on servers outside the user’s country, raising questions about data security and privacy. Jason warns that if the Chinese government exerts pressure on DeepSeek, the platform may have limited options due to its location and regulations.

In conclusion, as DeepSeek continues to make strides in the AI field, questions surrounding its data usage, compliance with regulations, and connections to sensitive issues persist, signaling a complex landscape at the intersection of technology and governance.